A chromatin subcompartment prediction method and system based on edge graph autoencoder

Abstract

The present invention provides a chromatin subcompartment prediction method and system based on an edge graph autoencoder, which relates to the technical field of chromatin subcompartment prediction, including: calculating the interaction strength according to first information, generating an interaction matrix by calculating the spatial proximity correlation strength between linear regions; constructing an interaction network according to the interaction matrix, and constructing interaction network graph structure data; performing marginalization information injection and anti-noise processing according to the interaction network graph structure data to generate a fusion feature representation; performing graph topology and semantic feature fusion according to the fusion feature representation to obtain a node deep embedding representation; generating a prediction result of the chromatin subcompartment according to the node deep embedding representation. Based on the edge graph autoencoder, the present invention realizes the unsupervised deep fusion of the chromatin interaction network topology structure and the node functional characteristics through the joint optimization of the marginalization information injection mechanism of randomly masking the node features and the graph neighborhood feature propagation, thereby improving the accuracy of subcompartment detection.

Description

Technical Field

[0001] The present invention relates to the technical field of chromatin subcompartment prediction, and in particular to a chromatin subcompartment prediction method and system based on an edge graph autoencoder. Background Art

[0002] Mammalian genomic DNA is highly folded into a complex three-dimensional conformation within the cell nucleus. Its dynamic folding pattern directly regulates core biological processes such as gene transcription, replication, and damage repair. The development of chromatin conformation capture (Hi-C) technology has provided a key tool for deciphering the three-dimensional structure of the genome. Based on Hi-C data, Lieberman-Aiden et al. first divided chromatin into open A compartments (active transcription regions) and closed B compartments (heterochromatin regions) when studying human chromatin interactions. However, this coarse-grained division failed to reveal the functional heterogeneity of subcompartments (such as A1 / A2 and B1-B4). Recent studies have confirmed that subcompartments are tissue-specific and disease-associated, making their precise detection crucial for understanding gene regulatory mechanisms. Existing methods for predicting chromatin subcompartments suffer from the following technical limitations: Unsupervised clustering methods based on epigenetic signatures, while capable of capturing local functional properties, ignore the topological structure of interaction networks, resulting in a loss of spatial proximity information; strategies relying on modularization of interaction networks, while capable of identifying spatial proximity relationships, fail to incorporate node functional characteristics, making it difficult to distinguish functionally heterogeneous subclasses; and algorithms employing supervised learning models, while performing well on specific datasets, rely on species- or tissue-specific marker data, limiting their ability to generalize across scenarios. Due to the limited use of features or flawed learning paradigms, these methods struggle to balance detection accuracy and generalizability, hindering in-depth research into the mechanisms regulating three-dimensional chromatin dynamics.

[0003] Based on the above shortcomings of the existing technology, there is an urgent need for a chromatin subcompartment prediction method and system based on edge graph autoencoder. Summary of the Invention

[0004] The present invention aims to provide a chromatin subcompartment prediction method based on an edge graph autoencoder to improve the above-mentioned problems. To achieve the above-mentioned object, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present application provides a chromatin subcompartment prediction method based on an edge graph autoencoder, comprising:

[0006] Acquiring first information, the first information including a chromatin three-dimensional conformation sequencing dataset and corresponding genome linear region division information;

[0007] Calculating the interaction strength based on the first information, and generating an interaction matrix by calculating the spatial proximity correlation strength between linear regions;

[0008] Constructing an interaction network based on the interaction matrix, determining significant interaction relationships between linear regions through statistical distribution thresholds, and constructing interaction network graph structure data;

[0009] According to the interaction network graph structure data, marginalization information injection and anti-noise processing are performed, noise perturbation features are generated by randomly masking some dimensional values ​​of node feature vectors, and marginalization topology information is fused in combination with the graph neighborhood feature propagation mechanism to generate a fused feature representation;

[0010] Performing graph topology and semantic feature fusion based on the fused feature representation to obtain a node deep embedding representation;

[0011] Based on the node deep embedding representation, the node similarity in the embedding space is divided through a spectral clustering framework to generate a prediction result of the chromatin subcompartment, which is a functional subcompartment category label corresponding to the linear region of the genome.

[0012] In a second aspect, the present application also provides a chromatin subcompartment prediction system based on an edge graph autoencoder, comprising:

[0013] An acquisition module is configured to acquire first information, wherein the first information includes a chromatin three-dimensional conformation sequencing dataset and corresponding genome linear region division information;

[0014] a calculation module, configured to calculate the interaction strength according to the first information, and generate an interaction matrix by calculating the spatial proximity correlation strength between linear regions;

[0015] A construction module is used to construct an interaction network according to the interaction matrix, determine the significant interaction relationship between linear regions through statistical distribution thresholds, and construct interaction network graph structure data;

[0016] A fusion module is used to perform marginalization information injection and anti-noise processing based on the interaction network graph structure data, generate noise perturbation features by randomly masking some dimensional values ​​of node feature vectors, and fuse marginalization topology information with the graph neighborhood feature propagation mechanism to generate a fused feature representation;

[0017] An extraction module, configured to fuse graph topology and semantic features based on the fused feature representation to obtain a deep embedding representation of the node;

[0018] A prediction module is used to divide the node similarities in the embedding space according to the node deep embedding representation through a spectral clustering framework to generate prediction results of chromatin subcompartments, wherein the prediction results are functional subcompartment category labels corresponding to the linear regions of the genome.

[0019] The beneficial effects of the present invention are:

[0020] Based on the edge graph autoencoder, the present invention breaks through the traditional algorithm's dependence on a single data source through the joint optimization of the marginalization information injection mechanism of randomly masking node features and the propagation of graph neighborhood features, and realizes the unsupervised deep fusion of the chromatin interaction network topology and node functional characteristics, significantly improving the accuracy of sub-compartment detection; utilizing the analytical solution characteristics of the graph autoencoder, the global optimal embedding representation can be directly solved without relying on iterative training, greatly reducing computational complexity and resource consumption; combining the sparse interaction network constructed with dynamic significance thresholds and the cross-scene spectral clustering framework, it effectively solves the problem of insufficient generalization caused by isolated feature utilization or reliance on supervised learning in existing methods, and provides technical support for revealing the three-dimensional conformational basis of chromatin for the formation of traits and diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 A schematic flow chart of a chromatin subcompartment prediction method based on an edge graph autoencoder according to an embodiment of the present invention;

[0023] Figure 2 Schematic diagram of the structure of a chromatin subcompartment prediction system based on an edge graph autoencoder according to an embodiment of the present invention;

[0024] Figure 3 This is the overall flowchart of the chromatin subcompartment prediction method based on edge graph autoencoder;

[0025] Figure 4 The enrichment index evaluation results of PEI with both ends located in the same type of subcompartment by the three algorithms;

[0026] Figure 5 The silhouette coefficient evaluation results of the three algorithms are shown in Figure 2.

[0027] Figure 6 The gene expression correlation fold ratio evaluation results of the three algorithms are shown in Figure 2.

[0028] Markings in the figure: 901, acquisition module; 902, calculation module; 903, construction module; 904, fusion module; 905, extraction module; 906, prediction module. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.

[0031] Example 1:

[0032] This embodiment provides a chromatin subcompartment prediction method based on an edge graph autoencoder.

[0033] See also Figure 1 , the figure shows that the method includes steps S100 to S600.

[0034] Step S100: obtaining first information, the first information including a chromatin three-dimensional conformation sequencing dataset and corresponding genome linear region division information;

[0035] Preferably, in this step, a chromatin three-dimensional conformation sequencing dataset generated based on Hi-C technology (e.g., K562 cell line data with index number GSE63525) is downloaded from a public database (such as GEO), and the genome is linearly divided into continuous regions of fixed length (bin units) according to an accuracy of 100 kb. This accuracy selection is consistent with existing algorithms and early subcompartment definition standards to ensure the standardization of data processing and the comparability of cross-research results.

[0036] Step S200: Calculate the interaction strength according to the first information, and generate an interaction matrix by calculating the spatial proximity correlation strength between linear regions;

[0037] As can be understood, based on the standardized genome partitioning information and raw Hi-C sequencing data obtained in step S100, this step normalizes the raw contact frequencies by eliminating sequencing technology bias, generating an interaction matrix representing the true spatial proximity strength between linear regions. This converts discrete sequencing reads into a continuous numerical matrix representation, providing quantitative input for the subsequent construction of the chromatin interaction network.

[0038] Step S300: constructing an interaction network based on the interaction matrix, determining the significant interaction relationship between linear regions through statistical distribution thresholds, and constructing interaction network graph structure data;

[0039] It should be noted that this step dynamically determines the significance threshold within different genomic distance intervals by analyzing the statistical distribution characteristics of the interaction strength, screens out the interaction relationships that represent the real three-dimensional space proximity, and eliminates random noise and non-specific signals; specifically, taking each linear region as a node, the node pairs whose interaction strength exceeds the corresponding distance interval threshold are defined as edges, and the node feature vectors are generated based on the distribution entropy and topological density of the interaction strength, and finally a sparse chromatin interaction network diagram is constructed.

[0040] Step S400: Perform marginalization information injection and anti-noise processing based on the interaction network graph structure data, generate noise perturbation features by randomly masking some dimensional values ​​of node feature vectors, and fuse marginalization topology information with the graph neighborhood feature propagation mechanism to generate a fused feature representation;

[0041] It can be understood that this step significantly improves the model's adaptability to sparse or low-quality Hi-C data through controllable noise injection and graph-structure-driven feature compensation, overcomes the overfitting or local optimality problems caused by the direct use of raw features in traditional algorithms, and provides highly generalizable input features for unsupervised clustering.

[0042] Step S500: fusion of graph topology and semantic features based on the fused feature representation to obtain a node deep embedding representation;

[0043] It should be noted that this step deeply optimizes the fused feature representation, extracting high-level semantic information from nodes and integrating global topological associations, ultimately generating a low-dimensional, dense deep embedding representation of the nodes. This process is achieved through an unsupervised approach, modeling complex spatial-functional associations without relying on labeled data. This overcomes the detection accuracy bottleneck caused by feature fragmentation in traditional methods and provides an embedding representation that combines local structural tightness with global functional consistency for subsequent spectral clustering.

[0044] Step S600: Based on the node deep embedding representation, the node similarity in the embedding space is divided using a spectral clustering framework to generate a prediction result of the chromatin subcompartment. The prediction result is a functional subcompartment category label corresponding to the genomic linear region.

[0045] It can be understood that this step constructs a similarity matrix that integrates functional association and genomic linear proximity, combines it with the spectral clustering algorithm to perform unsupervised division of chromatin regions, and finally outputs the functional sub-compartment category label (such as A1, A2, B1-B4) corresponding to each genomic linear region (100kbbin).

[0046] The overall flow chart of the present invention is as follows Figure 3 As shown in the figure, Even chrs / Odd chrs means grouping chromosomes according to their parity numbers; GAE means graph autoencoder.

[0047] Furthermore, step S200 includes steps S210 to S230.

[0048] Step S210: performing contact frequency correction processing according to the first information, by iteratively balancing the global coverage deviation of the matrix rows and columns, and preferentially retaining the three-dimensional spatial adjacent correlation signals to obtain a normalized contact frequency matrix;

[0049] This step normalizes the contact frequency matrix using an iterative balancing algorithm to eliminate technical biases such as uneven sequencing depth and differences in chromatin accessibility. Specifically, by iteratively adjusting the sum of the matrix rows and columns to a balanced state, the normalized contact frequency matrix is ​​generated, prioritizing the preservation of three-dimensional spatial proximity signals and suppressing systematic errors caused by experimental biases (such as PCR amplification bias and differences in enzyme digestion efficiency). If the iterations do not converge, coverage normalization is used instead to ensure stable output even with low-quality data.

[0050] Step S220: performing interaction signal extraction processing based on the normalized contact frequency matrix, separating the real space neighboring signals from the random noise by fitting the overdiscrete statistical distribution of the contact frequency between the linear regions, and obtaining an intensity matrix that removes nonspecific interactions;

[0051] This step distinguishes true spatial proximity signals from random noise through overdiscrete statistical distribution modeling. Since nonspecific interactions in Hi-C data follow a Poisson distribution, while true interactions exhibit overdiscrete characteristics due to the three-dimensional folding of chromatin, we fit the statistical distribution of contact frequencies between linear regions, calculate the significance p-value of each interaction strength, filter out nonspecific signals, and generate a denoised interaction strength matrix.

[0052] Step S230: quantify spatial correlation according to the interaction strength matrix, dynamically weight the interaction strength between linear regions through the nonlinear attenuation function of genomic distance, adapt the nonlinear mapping relationship between the three-dimensional chromatin folding pattern and the linear distance, and generate an interaction matrix.

[0053] This step dynamically weights interaction strengths using a nonlinear decay function based on genomic distance to quantify the correlation between 3D chromatin folding patterns and linear distance. Preferably, a piecewise decay function (e.g., short-range power-law decay, long-range exponential decay) is constructed to dynamically weight interaction strengths based on linear genomic distance: close-range interactions (<1Mb) are strengthened (reflecting localized chromatin looping), while long-range interactions (>10Mb) are weakened (to suppress random long-range contact noise).

[0054] Furthermore, step S300 includes steps S310 to S330.

[0055] Step S310: Perform significance distribution modeling based on the interaction matrix, calculate the extreme value distribution thresholds within different genomic distance intervals by fitting the long-tail statistical distribution characteristics of the interaction strength, and obtain the dynamic significance judgment threshold;

[0056] Specifically, this step divides the interactions between linear regions into multiple intervals according to genomic distance, independently models the interaction intensity distribution within each interval, determines the tail extreme value distribution parameters through maximum likelihood estimation, and calculates the dynamic threshold that meets the preset significance level.

[0057] Step S320: Screen the interaction relationships based on the dynamic significance threshold, retain the top K% high-intensity interactions through inter-partition threshold filtering and interaction strength sorting, and obtain a sparse interaction adjacency matrix with non-significant noise removed, where K% represents the proportion threshold;

[0058] It can be understood that this step retains interactions within each distance interval that exceed the dynamic threshold for that interval, while further truncating high-intensity interactions by a proportional threshold (e.g., K%, where K is adaptively adjusted based on interval sparsity) to generate a sparse interaction adjacency matrix. This strategy significantly reduces network redundancy while preserving key three-dimensional spatial interactions, providing high signal-to-noise ratio input for subsequent graph modeling.

[0059] Step S330: Graph structure data generation processing is performed based on the sparse interaction adjacency matrix, and node feature vectors are calculated jointly by interaction intensity distribution entropy and topological closeness to construct interaction network graph structure data.

[0060] Among them, distribution entropy is expressed as: by calculating the Shannon entropy of the distribution of the interaction intensity between a node and all other nodes, the diversity of its interaction pattern is quantified (high entropy indicates extensive connection, low entropy indicates specific connection); topological closeness is expressed as: based on the betweenness centrality or clustering coefficient of the node in the graph, its topological importance in the network is measured (such as hub node or local cluster core).

[0061] Furthermore, step S400 includes steps S410 to S430.

[0062] Step S410: Perform random noise injection processing based on the interaction network graph structure data, and simulate the sequencing data missing scenario by randomly masking some dimensions of the node feature vector to generate noise disturbance features that are resistant to overfitting;

[0063] Preferably, this step randomly deletes 20% of the dimensions in each node's feature vector (this ratio was optimized and determined through multiple comparative experiments) to simulate data loss scenarios caused by technical limitations in Hi-C sequencing, generating noise-perturbed features that are resistant to overfitting. Specifically, a Bernoulli distribution is used to independently sample the feature dimensions of each node (with a 20% masking probability). This forces the model to learn robust representations of key functional features during training, preventing overfitting to noisy or sparse data.

[0064] Step S420: Perform marginal information fusion processing based on the noise disturbance features, aggregate the interaction relationship features of adjacent nodes through the neighborhood feature propagation mechanism of the graph autoencoder, and generate topology enhancement features;

[0065] It can be understood that this step aggregates the interaction relationship features of adjacent nodes to compensate for the masking loss through the neighborhood feature propagation mechanism of the graph autoencoder. Specifically, the features of the first-order neighborhood nodes of each node are weighted averaged (the weight is determined by the interaction strength) to generate topological enhancement features. For example, if a feature dimension of node A is masked, but its neighboring nodes B / C have high interaction strength in this dimension, the semantic information of this dimension is restored through neighborhood propagation. The 20% masking ratio has been experimentally verified to balance the weight of noise introduction and information retention, avoiding excessive damage to the original features while ensuring that the model fully relies on the topological structure for feature compensation, thereby improving adaptability to low-coverage data.

[0066] Step S430: perform anti-noise joint optimization based on the topology enhancement features, perform anti-noise feature joint processing, combine the original node features with the topology fusion features through linear weighting, and generate a fusion feature representation.

[0067] This step performs a linear weighted fusion of the original node features and the topology-enhanced features to generate the final fused feature representation. Residual connections are used to balance the deviations between feature masking and topology compensation, ensuring that the model fully integrates network structure information while preserving the original functional properties.

[0068] Furthermore, step S500 includes steps S510 to S530.

[0069] Step S510: extract local interaction patterns based on the fused feature representation, aggregate the node's own features and the weighted features of its first-order neighboring nodes through a graph convolutional network to generate a primary embedding representation;

[0070] Specifically, a weighted average of the features of each node's first-order neighborhood (directly connected nodes) is performed, with the weights determined by the strength of the interactions, thereby capturing the local topological correlations formed by chromatin folding. This operation simulates the physical proximity effects of chromatin at the nanoscale, providing foundational features for subsequent modeling of long-range functional associations and resolving the problem of blurred subcompartmental boundaries caused by traditional methods (such as hidden Markov models) that ignore spatial proximity.

[0071] Step S520: Perform global semantic association modeling based on the primary embedding representation, capture cross-region long-range interaction dependencies through a stacked multi-head graph attention mechanism, and generate an intermediate embedding representation;

[0072] Based on the primary embedding representation, this step uses a stacked multi-head graph attention mechanism (multi-head GAT) to model long-range cross-regional interactions and dependencies, generating a mid-level embedding representation. Each attention head independently learns functional association weights between nodes (e.g., based on gene co-expression or epigenetic similarity), focusing on distant nodes that are not directly connected but functionally synergistic (e.g., enhancer-promoter loops across chromosomes). The multi-head mechanism enhances the model's ability to represent complex three-dimensional functional networks by concurrently capturing association patterns across different semantic spaces (e.g., transcription factor binding preferences, chromatin accessibility matching).

[0073] Step S530: Perform nonlinear fusion of high-order features based on the intermediate embedding representation, integrate multi-order topological semantic information through residual connections, and generate a deep embedding representation of the node.

[0074] Specifically, this step connects the primary embedding (local topology), the intermediate embedding (global functionality), and the original fused features across layers, achieving high-order feature fusion through nonlinear activation functions, preserving multi-granular semantic information (such as local compactness and global functional modularity). This strategy effectively alleviates the gradient decay problem in deep network training and improves the adaptability of the embedding representation to the multi-scale structure of chromatin, ensuring that the subcompartmental partitioning results meet both spatial proximity and functional consistency.

[0075] Furthermore, step S600 includes steps S610 to S630.

[0076] Step S610: construct a spectral similarity matrix based on the node deep embedding representation, and generate a spectral similarity matrix that integrates functional association and linear proximity by jointly weighting the Gaussian kernel function and the genomic linear position attenuation factor;

[0077] This step constructs a jointly weighted spectral similarity matrix by fusing functional association and genomic linear proximity, which can be expressed mathematically as follows:

[0078] ;

[0079] Among them, W ij represents the joint weighted similarity of nodes i and j; z i , z j represents the deep embedding vector of nodes i and j; σ represents the Gaussian kernel bandwidth parameter, which is usually 1.5 times the standard deviation of the embedding vector; d ij represents the physical distance between nodes i and j on the linear sequence of the genome; L decay is the decay length constant, which is set to 1Mb by default.

[0080] Step S620: performing low-dimensional spectrum embedding mapping according to the spectrum similarity matrix, extracting eigenvectors by eigendecomposition of the normalized Laplace matrix, and generating a low-dimensional spectrum embedding space;

[0081] The expression is:

[0082] ;

[0083] Where L represents the normalized symmetric Laplace matrix with dimension n×n (n is the total number of nodes); D represents the degree matrix; 1 represents the all-1 vector; W represents the spectral similarity matrix, which is the same as W ij consistent.

[0084] Step S630: Clustering and verification processing is performed based on the low-dimensional spectral embedding space. Subcompartment labels are generated using the K-means algorithm combined with silhouette coefficient maximization. The rationality of the compartment division is verified by promoter-enhancer interaction enrichment and gene expression correlation.

[0085] First, the silhouette coefficient is a mathematical indicator used to quantify the closeness and separation of clusters. The formula is:

[0086] ;

[0087] Among them, s(m) represents the silhouette coefficient of sample m; a(m) represents the average distance from sample m to other samples in the same cluster, reflecting the compactness within the cluster (the smaller the better); b(m) represents the average distance from sample m to all samples in the nearest other cluster, reflecting the separation between clusters (the larger the better).

[0088] Secondly, the proportion of promoter-enhancer interactions (PEIs) in subcompartments of the same type is used as another independent evaluation index. Preferably, ChIA-PET (capturing promoter-enhancer interactions (PEIs)) data of PolII (RNA polymerase II) is downloaded from a public database to evaluate compartmental detection. Theoretically, the number of promoter-enhancer interactions (PEIs) in subcompartments of the same type should be significantly higher than that in subcompartments of different types, and a more significant difference between the two indicates better subcompartmental detection.

[0089] Next, the functional consistency of the subcompartments was verified based on the gene expression correlation difference data. By calculating the Spearman correlation coefficient of the gene expression within and across subcompartments, it was found that the correlation between gene expression in the same type of subcompartments was significantly higher than the correlation between genes in different types of subcompartments. The more significant the difference between the two, the better the subcompartment detection effect.

[0090] Furthermore, in the result verification, by comparing the image data of the present invention with the existing method (such as Figures 4 to 6 ), we can intuitively observe the significant advantages of this method in multi-dimensional indicators, where Deeplearning_m1 represents this method, algorithm A represents the hidden Markov model (HMM), and algorithm B represents the sub-compartment identification algorithm (Sub-Compartment Identifier, SCI). Figure 4 The enrichment of functional interactions (PEIs) within the same subcompartment using different algorithms is shown. The correlation ratio, a measure of consistency between chromatin subcompartmental delineation and actual biological function, shows that this method significantly outperforms other methods in terms of histogram height and error margin, demonstrating its stability and accuracy in distinguishing functionally associated regions. This result is directly related to the design of the joint function-space modeling in step S610, which uses a dynamic significance threshold to screen key interaction signals, effectively preventing noise from interfering with functional associations.

[0091] according to Figure 5 The silhouette coefficient is compared and analyzed in the histogram. Silhouette_score represents the silhouette coefficient. The silhouette coefficient of this method (Deeplearning_m1) is positive and the highest, significantly outperforming algorithms B and C. This intuitively shows that its clustering results are ahead in terms of compactness (the degree of aggregation of similar samples) and separation (the ability to distinguish heterogeneous samples).

[0092] according to Figure 6 A histogram analysis of gene correlation ratios as a function of distance is presented. Distance between genes represents the linear genomic distance between gene pairs. Our method (Deeplearning_m1) maintains a significant advantage across the entire distance range of 0.5-9.5 Mb. In particular, at long distances (>5.5 Mb), the correlation ratio exceeds 3.0, while the ratios for algorithms B and C in this range are less than 2.0. This result demonstrates that our method can more accurately capture long-range functional associations (such as enhancer-promoter interactions across TADs), overcoming the long-range signal loss problem often associated with traditional methods due to linear distance decay.

[0093] The correlation between the above image results and the technical process shows that this method has achieved breakthrough improvements in both mathematical indicators and biological significance by integrating topological and functional features, optimizing noise suppression strategies and introducing multi-dimensional verification mechanisms, providing reliable technical support for the analysis of chromatin three-dimensional conformation.

[0094] Example 2:

[0095] like Figure 2 As shown, this embodiment provides a chromatin subcompartment prediction system based on an edge graph autoencoder, the system comprising:

[0096] An acquisition module 901 is configured to acquire first information, the first information including a chromatin three-dimensional conformation sequencing dataset and corresponding genome linear region division information;

[0097] A calculation module 902 is configured to calculate the interaction strength according to the first information, and generate an interaction matrix by calculating the spatial proximity correlation strength between linear regions;

[0098] Construction module 903, for constructing an interaction network based on the interaction matrix, determining significant interaction relationships between linear regions through statistical distribution thresholds, and constructing interaction network graph structure data;

[0099] Fusion module 904 is used to perform marginalization information injection and anti-noise processing based on the interaction network graph structure data, generate noise perturbation features by randomly masking some dimensional values ​​of node feature vectors, and fuse marginalization topology information with the graph neighborhood feature propagation mechanism to generate a fused feature representation;

[0100] Extraction module 905, used to fuse graph topology and semantic features based on the fused feature representation to obtain a deep embedding representation of the node;

[0101] The prediction module 906 is used to divide the node similarities in the embedding space according to the node deep embedding representation through the spectral clustering framework to generate the prediction results of the chromatin subcompartment, and the prediction results are the functional subcompartment category labels corresponding to the linear regions of the genome.

[0102] In some embodiments disclosed in this application, the calculation module 902 includes:

[0103] a first calculation unit, configured to perform contact frequency correction processing according to the first information, by iteratively balancing global coverage deviations of matrix rows and columns, and preferentially retaining adjacent correlation signals in three-dimensional space, to obtain a normalized contact frequency matrix;

[0104] The second calculation unit is used to extract and process the interaction signal based on the normalized contact frequency matrix. By fitting the over-discrete statistical distribution of the contact frequency between linear regions, the real space neighboring signal and the random noise are separated to obtain the intensity matrix that removes non-specific interactions.

[0105] The third calculation unit is used to quantify spatial correlation based on the interaction strength matrix, dynamically weight the interaction strength between linear regions through the nonlinear attenuation function of genomic distance, adapt the nonlinear mapping relationship between the three-dimensional chromatin folding pattern and linear distance, and generate an interaction matrix.

[0106] In some embodiments disclosed in this application, the building block 903 includes:

[0107] The first construction unit is used to perform significance distribution modeling based on the interaction matrix, calculate the extreme value distribution threshold within different genomic distance intervals by fitting the long-tail statistical distribution characteristics of the interaction intensity, and obtain the dynamic significance judgment threshold;

[0108] The second construction unit is used to screen the interaction relationships according to the dynamic significance judgment threshold. The top K% high-intensity interactions are retained through interval threshold filtering and interaction strength sorting to obtain a sparse interaction adjacency matrix that removes non-significant noise, where K% represents the proportion threshold.

[0109] The third construction unit is used to generate and process graph structure data based on the sparse interaction adjacency matrix, and to construct the interaction network graph structure data by jointly calculating the node feature vector through the interaction intensity distribution entropy and topological closeness.

[0110] In some embodiments disclosed herein, the fusion module 904 includes:

[0111] The first fusion unit is used to perform random noise injection processing based on the interaction network graph structure data, and to simulate the sequencing data missing scenario by randomly masking some dimensions of the node feature vector to generate noise perturbation features that are resistant to overfitting;

[0112] The second fusion unit is used to perform marginal information fusion processing based on the noise perturbation characteristics, aggregate the interaction relationship features of adjacent nodes through the neighborhood feature propagation mechanism of the graph autoencoder, and generate topology enhancement features;

[0113] The third fusion unit is used to perform anti-noise joint optimization based on the topology enhancement features, perform anti-noise feature joint processing, combine the original node features with the topology fusion features through linear weighting, and generate a fusion feature representation.

[0114] In some embodiments disclosed in this application, the extraction module 905 includes:

[0115] The first extraction unit is used to perform random noise injection processing based on the interaction network graph structure data, and to simulate the sequencing data missing scenario by randomly masking some dimensions of the node feature vector to generate noise disturbance features that are resistant to overfitting;

[0116] The second extraction unit is used to extract local interaction patterns based on the fused feature representation, and to generate a primary embedding representation by aggregating the node's own features and the weighted features of its first-order neighboring nodes through a graph convolutional network;

[0117] The third extraction unit is used to perform global semantic association modeling based on the primary embedding representation, capture cross-region long-range interaction dependencies through the stacked multi-head graph attention mechanism, and generate an intermediate embedding representation;

[0118] The fourth extraction unit is used to perform nonlinear fusion of high-order features based on the intermediate embedding representation, integrate multi-order topological semantic information through residual connections, and generate a deep embedding representation of the node.

[0119] In some embodiments disclosed herein, the prediction module 906 includes:

[0120] The first prediction unit is used to construct a spectral similarity matrix based on the node deep embedding representation, and generate a spectral similarity matrix that integrates functional association and linear proximity by jointly weighting the Gaussian kernel function and the genomic linear position attenuation factor;

[0121] The second prediction unit is used to perform low-dimensional spectrum embedding mapping according to the spectrum similarity matrix, extract feature vectors by normalized Laplace matrix eigendecomposition, and generate a low-dimensional spectrum embedding space;

[0122] The third prediction unit is used to perform clustering and verification processing based on the low-dimensional spectral embedding space. The subcompartment labels are generated by combining the K-means algorithm with the silhouette coefficient maximization, and the rationality of the compartment division is verified by the promoter-enhancer interaction enrichment and gene expression correlation.

[0123] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A chromatin subcompartment prediction method based on edge graph autoencoder, characterized in that: include: Acquiring first information, the first information including a chromatin three-dimensional conformation sequencing dataset and corresponding genome linear region division information; Calculating the interaction strength based on the first information, and generating an interaction matrix by calculating the spatial proximity correlation strength between linear regions; Constructing an interaction network based on the interaction matrix, determining significant interaction relationships between linear regions through statistical distribution thresholds, and constructing interaction network graph structure data; According to the interaction network graph structure data, marginal information injection and anti-noise processing are performed, noise perturbation features are generated by randomly masking some dimensional values ​​of node feature vectors, and marginal topological information is fused in combination with the graph neighborhood feature propagation mechanism to generate a fused feature representation; Performing graph topology and semantic feature fusion based on the fused feature representation to obtain a node deep embedding representation; Based on the node deep embedding representation, the node similarity in the embedding space is divided through a spectral clustering framework to generate a prediction result of the chromatin subcompartment, which is a functional subcompartment category label corresponding to the linear region of the genome.

2. The chromatin subcompartment prediction method based on edge graph autoencoder according to claim 1, characterized in that Calculating the interaction strength according to the first information and generating an interaction matrix by calculating the spatial proximity correlation strength between linear regions includes: performing contact frequency correction processing according to the first information, by iteratively balancing the global coverage deviation of the matrix rows and columns, and preferentially retaining the three-dimensional spatial neighboring correlation signals to obtain a normalized contact frequency matrix; Performing interaction signal extraction processing according to the normalized contact frequency matrix, separating real space neighboring signals from random noise by fitting the overdiscrete statistical distribution of contact frequencies between linear regions, and obtaining an intensity matrix that removes nonspecific interactions; The spatial correlation is quantified according to the interaction strength matrix, the interaction strength between linear regions is dynamically weighted by the nonlinear attenuation function of genomic distance, the nonlinear mapping relationship between the three-dimensional chromatin folding pattern and the linear distance is adapted, and the interaction matrix is ​​generated.

3. The chromatin subcompartment prediction method based on edge graph autoencoder according to claim 1, characterized in that An interaction network is constructed according to the interaction matrix, and significant interaction relationships between linear regions are determined by statistical distribution thresholds to construct interaction network graph structure data, including: Performing significance distribution modeling based on the interaction matrix, and calculating extreme value distribution thresholds within different genomic distance intervals by fitting the long-tail statistical distribution characteristics of the interaction intensity, thereby obtaining a dynamic significance determination threshold; Interaction relationships are screened according to the dynamic significance judgment threshold, and the top K% high-intensity interactions are retained through interval threshold filtering and interaction strength sorting to obtain a sparse interaction adjacency matrix that removes non-significant noise, where K% represents the proportion threshold; Graph structure data generation processing is performed based on the sparse interaction adjacency matrix, and node feature vectors are calculated jointly by interaction intensity distribution entropy and topological closeness to construct interaction network graph structure data.

4. The chromatin subcompartment prediction method based on edge graph autoencoder according to claim 1, characterized in that According to the interaction network graph structure data, marginal information injection and anti-noise processing are performed, noise perturbation features are generated by randomly masking some dimensional values ​​of node feature vectors, and marginal topological information is fused in combination with the graph neighborhood feature propagation mechanism to generate a fused feature representation, including: Performing random noise injection processing on the interaction network graph structure data, numerically simulating a sequencing data missing scenario by randomly masking some dimensions of the node feature vector, and generating noise perturbation features that are resistant to overfitting; Perform marginal information fusion processing based on the noise perturbation features, aggregate the interaction relationship features of adjacent nodes through the neighborhood feature propagation mechanism of the graph autoencoder, and generate topology enhancement features; Anti-noise joint optimization is performed based on the topology enhancement features, anti-noise feature joint processing is performed, and the original node features are combined with the topology fusion features through linear weighting to generate a fusion feature representation.

5. The chromatin subcompartment prediction method based on edge graph autoencoder according to claim 1, characterized in that Graph topology and semantic features are fused according to the fused feature representation to obtain a deep embedding representation of the node, including: Extracting local interaction patterns based on the fused feature representation, aggregating node features and weighted features of its first-order neighboring nodes through a graph convolutional network to generate a primary embedding representation; Based on the primary embedding representation, global semantic association modeling is performed, and cross-region long-range interaction dependencies are captured through a stacked multi-head graph attention mechanism to generate an intermediate embedding representation; High-order features are nonlinearly fused based on the intermediate embedding representation, and multi-order topological semantic information is integrated through residual connections to generate a deep embedding representation of the node.

6. A chromatin subcompartment prediction system based on edge graph autoencoder, characterized in that include: An acquisition module is configured to acquire first information, wherein the first information includes a chromatin three-dimensional conformation sequencing dataset and corresponding genome linear region division information; a calculation module, configured to calculate the interaction strength according to the first information, and generate an interaction matrix by calculating the spatial proximity correlation strength between linear regions; A construction module is used to construct an interaction network according to the interaction matrix, determine the significant interaction relationship between linear regions through statistical distribution thresholds, and construct interaction network graph structure data; A fusion module is used to perform marginalization information injection and anti-noise processing based on the interaction network graph structure data, generate noise perturbation features by randomly masking some dimensional values ​​of node feature vectors, and fuse marginalization topology information with the graph neighborhood feature propagation mechanism to generate a fused feature representation; An extraction module, configured to fuse graph topology and semantic features based on the fused feature representation to obtain a deep embedding representation of the node; A prediction module is used to divide the node similarities in the embedding space according to the node deep embedding representation through a spectral clustering framework to generate prediction results of chromatin subcompartments, wherein the prediction results are functional subcompartment category labels corresponding to the linear regions of the genome.

7. The chromatin subcompartment prediction system based on edge graph autoencoder according to claim 6, characterized in that The calculation module includes: a first calculation unit, configured to perform contact frequency correction processing according to the first information, by iteratively balancing global coverage deviations of matrix rows and columns, and preferentially retaining adjacent correlation signals in three-dimensional space, to obtain a normalized contact frequency matrix; a second computing unit configured to perform interaction signal extraction processing based on the normalized contact frequency matrix, separate the real space neighboring signals from the random noise by fitting the overdiscrete statistical distribution of the contact frequency between the linear regions, and obtain an intensity matrix that removes non-specific interactions; The third calculation unit is used to quantify spatial correlation according to the interaction strength matrix, dynamically weight the interaction strength between linear regions through the nonlinear attenuation function of genomic distance, adapt the nonlinear mapping relationship between the three-dimensional chromatin folding pattern and the linear distance, and generate an interaction matrix.

8. The chromatin subcompartment prediction system based on edge graph autoencoder according to claim 6, characterized in that: The building blocks include: The first construction unit is used to perform significance distribution modeling based on the interaction matrix, calculate the extreme value distribution threshold in different genomic distance intervals by fitting the long-tail statistical distribution characteristics of the interaction intensity, and obtain a dynamic significance judgment threshold; The second construction unit is used to screen the interaction relationship according to the dynamic significance judgment threshold, retain the top K% high-intensity interactions through interval threshold filtering and interaction strength sorting, and obtain a sparse interaction adjacency matrix with non-significant noise removed, where K% represents the proportion threshold; The third construction unit is used to generate and process graph structure data according to the sparse interaction adjacency matrix, and construct the interaction network graph structure data by jointly calculating the node feature vector by the interaction intensity distribution entropy and the topological closeness.

9. The chromatin subcompartment prediction system based on edge graph autoencoder according to claim 6, characterized in that The fusion module includes: A first fusion unit is configured to perform random noise injection processing based on the interaction network graph structure data, numerically simulate a sequencing data missing scenario by randomly masking some dimensions of the node feature vector, and generate noise disturbance features that are resistant to overfitting; A second fusion unit is configured to perform marginal information fusion processing based on the noise disturbance features, aggregate the interaction relationship features of adjacent nodes through the neighborhood feature propagation mechanism of the graph autoencoder, and generate topology enhancement features; The third fusion unit is used to perform anti-noise joint optimization based on the topology enhancement features, perform anti-noise feature joint processing, combine the original node features with the topology fusion features through linear weighting, and generate a fusion feature representation.

10. The chromatin subcompartment prediction system based on edge graph autoencoder according to claim 6, characterized in that: The extraction module includes: A first extraction unit is configured to perform random noise injection processing based on the interaction network graph structure data, numerically simulate a sequencing data missing scenario by randomly masking some dimensions of the node feature vector, and generate noise disturbance features that are resistant to overfitting; A second extraction unit is configured to extract local interaction patterns based on the fused feature representation, aggregate node features and weighted features of its first-order neighboring nodes through a graph convolutional network, and generate a primary embedding representation; A third extraction unit is configured to perform global semantic association modeling based on the primary embedding representation, capture cross-regional long-range interaction dependencies through a stacked multi-head graph attention mechanism, and generate an intermediate embedding representation; The fourth extraction unit is used to perform nonlinear fusion of high-order features based on the intermediate embedding representation, integrate multi-order topological semantic information through residual connections, and generate a node deep embedding representation.

Images